2025-05-02
After losing, Kasparov didn’t give up — he adapted.
He introduced “centaur chess”, pairing human intuition with machine power.
Kasparov’s idea of man + machine wasn’t just for chess.
This study brings that question into finance:
Can human analysts and AI models compete — or collaborate — in forecasting the market?
To explore this, the authors created an AI analyst and compared it to real professionals.
This study addresses three core questions:
Can an AI analyst predict 12-month-ahead stock returns more accurately than human analysts?
When do human analysts outperform AI — and why?
Does combining AI and human forecasts improve accuracy and reduce large errors?
In this study, both human and AI analysts forecast 12-month-ahead stock returns using target prices.
Target prices reflect analysts’ valuation beliefs
Predicted return = (Target – Current Price) / Current Price
Target prices are preferred over earnings forecasts:
The AI analyst is an ensemble of three models — combining strengths across ML and deep learning.
Random Forest
Gradient Boosting
LSTM (Long Short-Term Memory)
Final forecast = median prediction across all three models
Forecasts are made using a rolling 3-year window of past data
→ Simulates real-time prediction, avoiding lookahead bias
The AI model uses only public information
→ No human analyst inputs, no insider data
Data sources:
Over 1.15 million target price forecasts, covering thousands of U.S. firms
Coverage spans all major sectors, firm sizes, and time periods (2001–2018)
The AI analyst is trained on a wide variety of publicly available information.
The model draws from six key categories of inputs:
Firm Data: - Size, book-to-market, ROA - Leverage, accruals - Past returns (1–36 months), volatility - Amihud illiquidity
Industry Data: - Fama-French industry groupings - Industry competition (text-based) - Product market fluidity
Macro Data: - GDP growth, CPI, oil prices - Treasury yields, credit spreads - Market-level illiquidity
Filings (10-K, 10-Q, 8-K): - Sentiment and tone (Loughran–McDonald) - Readability and complexity - Text similarity and novelty
Sentiment: - RavenPack news sentiment - Twitter-based firm sentiment (Cao et al. 2021a)
Innovation: - Patent value estimates (Kogan et al. 2017)
AI uses only public data, with no access to analyst forecasts or private information.
While the AI model performs better overall, human analysts outperform in specific contexts:
Combining AI forecasts with analyst predictions and profiles creates a hybrid “centaur” analyst.
Extreme errors = forecasts in the top 10% of squared prediction errors
Man + Machine avoids:
Model Interpretability: Use SHAP values (SHapley Additive exPlanations) or partial dependence plots to explain predictions.
Robustness Across Market Conditions: Assess performance across macroeconomic regimes (e.g., recessions vs. expansions) to test model stability under systemic market stress.